A social media NMT engine for a low-resource language combination

© 2023 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC-BY-ND. A social media NMT engine for a low-resource language combination María do Campo Bayón Universitat Autònoma de Barcelona Spain maria.docampo@autonoma.cat Pilar Sánchez-Gijón Universitat Autònoma de Barcelona Spain pilar.sanchez.gijon@uab.cat Abstract The aim of this article is to present a new Neural Machine Translation (NMT) from Spanish into Galician for the social media domain that was trained with a Twitter cor- pus. Our main goal is to outline the methods used to build the corpus and the steps taken to train the engine in a low-resource lan- guage context. We evaluated the engine per- formance both with regular automatic met- rics and with a new methodology based on the non-inferiority process and contrasted this information with a human evaluation based on an error classification conducted by professional linguists. We will present the steps carried out following the conclusions of a previous pilot study, describe the new process followed, analyze the new engine and present the final conclusions. 1 Introduction In recent years, the low-resource languages domain has received some attention from our research community. Many papers covered different strategies to overcome the need for data to train engines for low-resource languages. Ranathunga et al. (2023) gave a complete overview of the main techniques and solutions employed in this field: data augmentation techniques, such as word or phrase replacement, back-translation, parallel data mining; unsupervised NMT; semi-supervised NMT; multilingual NMT; transfer learning in NMT; and zero-shot NMT. A considerable amount of work has been also done in social media research, mainly in sentiment analysis and translation of user-generated content fields. The majority of these papers are focused on Phrase-Based Machine Translation (PBMT) engines. Only Lohar et al. (2019) attempted to compare the machine translations of tweets using phrase-based and neural MT and the usage of different amounts and types of training corpora for each of the two approaches. The results of their research showed that using a tiny Twitter corpus is useless for NMT train- ing, although the system improved when using back- translation and out-of-domain corpora. This particu- lar procedure is the one used in our NMT training and adapted to a low-resource language combina- tion, from Spanish to Galician. 2 Background This contribution presents the most significant findings from a doctoral thesis on low-resource languages and NMT as a means of promoting and using a minority language in the context of social media1. It is carried out in the DespiteMT project framework, dedicated to researching the uses of MT applied to the media2. This study is based on a previous pilot completed in 2022, which focused on creating a Spanish into Galician NMT engine for social media and proposing a new methodology for evaluating this type of NMT engine (do Campo et al., 2022). For the pilot study, we created an NMT engine based on Joey through the online platform Mut- NMT3 (a minimalist NMT toolkit for novices, https://aclanthology.org/D19-3019/). The corpus used to …

Content truncated to 500 words